KR20220104722A - Workload-Based Clock Adjustment in Processing Units - Google Patents

Abstract

A graphics processing unit (GPU) 102 adjusts the frequency of the clock based on identifying the program threads 104 , 106 executing in the processing unit, and the program threads are detected based on the workload 116 , 117 to be executed. do. By adjusting the clock frequency based on the identified program thread, the processing unit adapts to the different processing needs of different program threads. Also, by identifying a program thread based on a workload, the processing unit conserves processing resources by adapting the clock frequency based on processing demands.

Description

Workload-Based Clock Adjustment in Processing Units

Processing systems often use specialized processing units to perform the specific operations for which the processing units are designed. For example, a processing system may utilize a graphics processing unit (GPU) to perform graphics and vector processing operations for the processing system. In some cases, the processing unit concurrently performs operations on behalf of different programs executing on the processing system. For example, the processing system may implement a virtualized computing environment, wherein the processing system concurrently executes multiple virtual machines (VMs) in one or more central processing units (CPUs). Each of the VMs requests the GPU of the processing system to execute graphics or vector processing operations, so the GPU is tasked with executing the operations concurrently on behalf of the different VMs. However, different programs may have different requirements, such as different power requirements, different maximum clock frequency requirements, and the like. Different requirements impose different processing demands on the GPU (or other processing unit), thereby negatively impacting the processing system as a whole.

BRIEF DESCRIPTION OF THE DRAWINGS The present disclosure may be better understood by those skilled in the art by reference to the accompanying drawings, and many features and advantages may become apparent to those skilled in the art. The use of the same reference symbols in different drawings indicates similar or identical items.
1 is a block diagram of a processing system including a graphics processing unit (GPU) that adjusts a clock frequency based on detection of an executing program thread in accordance with some embodiments.
FIG. 2 is a diagram illustrating an example of the GPU of FIG. 1 adjusting a clock frequency based on changes in the workload of the GPU in accordance with some embodiments.
3 is a block diagram of a clock frequency adjustment module of the GPU of FIG. 1 in accordance with some embodiments.
4 is a flow diagram of a method of adjusting a clock frequency of a processing unit based on detecting an executing program thread in accordance with some embodiments.

1-4 illustrate techniques for adjusting the frequency of a clock in a graphics processing unit (GPU) based on identifying a program thread executing in the central processing unit (CPU), the program thread based on the workload to be executed; detected. By adjusting the clock frequency based on the identified program thread, the processing unit adapts to the different processing needs of concurrently executing programs. Also, by identifying a program thread based on a workload, the processing unit conserves processing resources by adapting the clock frequency based on processing demands.

To illustrate by way of example, in some embodiments, the CPU executes two different programs (eg, two different virtual machines) concurrently. One of the programs designated as Program1 is, for example, a program required to execute operations at a relatively low frequency to maintain compatibility with other programs or systems. Another program, designated Program2, is the program required to execute operations at a relatively high frequency to meet performance targets, at least in some circumstances. Typically, the clock of the GPU executing the operations for Program1 and Program2 simultaneously is set higher among the different clock frequencies, regardless of which program's operations are being executed on the GPU, so that additional software or hardware for Program1 is not It requires compliance with its compatibility requirements. Using the techniques herein, the GPU detects which of Program1 and Program2 is being executed, and whether it is desirable to adjust the clock frequency for execution of the processing demands of the executing program.

Program to meet its processing requirements, and adjust the clock frequency of the GPU accordingly. The GPU thus meets the processing requirements for each of Program1 and Program2, and performs dynamically according to a changing workload, conserving processing resources.

Also, by changing the clock frequency based on the detected workload, and not solely by static conditions such as a program identifier value or specified time periods, the GPU can process processing for each program based on the needs of a given workload. specifications can be met, thereby saving processor resources. For example, in some cases, a higher-performance program may present a relatively light workload to the GPU, such that processing features may have a lower clock frequency. Using the techniques herein, the GPU maintains the frequency of the clock signal at a lower frequency under relatively light workloads, even if the corresponding program is associated with a relatively high specific clock frequency, so that they need to meet the program specification. Saves GPU resources when not present.

Referring to the drawings, FIG. 1 shows a block diagram of a processing system 100 including a central processing unit (CPU) 101 and a graphics processing unit (GPU) 102 in accordance with some embodiments. Processing system 100 is generally configured to execute instructions, which are organized as computer programs to perform tasks on behalf of an electronic device. Thus, in different embodiments, processing system 100 is part of an electronic device, such as a desktop or laptop computer, server, smartphone, tablet, game console, or the like. The processing system 100 includes additional components and modules not illustrated in FIG. 1 . For example, in some embodiments, the processing system includes one or more memory controllers, input/output controllers, network interfaces, etc. to perform tasks on behalf of the electronic device.

CPU 101 is generally configured to concurrently execute multiple programs and corresponding program threads. As used herein, a program thread refers to either an individual program (eg, an operating system, an application program, etc.) or an individual thread in a multithreaded program. In the example shown, CPU 101 executes two program threads, a designated program 103 and program 104 simultaneously. It will be appreciated, however, that the techniques described with respect to FIG. 1 are, in other embodiments, implemented in a processing system that executes N programs concurrently, where N is an integer greater than one. Thus, in some embodiments, CPU 101 implements a virtualized computing environment by concurrently executing multiple virtual machines, where programs 103 and 104 correspond to programs executed by different virtual machines. do. For example, in some embodiments, program 103 is an operating system associated with one virtual machine and program 104 is an operating system associated with a different virtual machine running on a processing system. For illustrative purposes, it is assumed that each program 103 and 104 has different processing specifications, such as different specific processing speeds, power consumption specifications, and the like. For example, in some embodiments, program 103 is a “legacy” program that is specified to run at a relatively low frequency to provide backward compatibility for other programs or hardware components of the processing system. In contrast, program 104 is a new program that is specified to run at a relatively high frequency to meet performance goals. As further described herein, GPU 102 may adjust specified parameters, and in particular clock frequency of GPU 102 , such that programs 103 and 104 each comply with their processing specifications. .

GPU 102 is generally configured to perform graphics and vector processing operations for a processing system. To illustrate, in some embodiments, in the course of executing programs 103 and 104 , CPU 101 may use GPU commands herein to request GPU 102 to perform specified operations. For example, create specific instructions referred to as instructions 105 and 107). Examples of GPU commands include draw commands that request the GPU to draw a specified object for display, vector commands that request the GPU to perform a specified vector operation, etc. . One or more CPUs provide GPU instructions to GPU 102 for execution. Each instruction is issued by and associated with a corresponding one of programs 103 and 104 . Thus, for example, in some embodiments, instruction 105 is a draw instruction that requests GPU 102 to draw one or more objects on behalf of program 103 , and instruction 107 is GPU 102 . is a draw command that requests to draw one or more objects on behalf of the program 104 .

It will be appreciated that CPU 101 executes programs 103 and 104 concurrently. Accordingly, in some embodiments, CPU 101 provides different instructions associated with different ones of programs 103 and 104 to GPU 102 in a time multiplexed manner for execution. To illustrate, in some embodiments, CPU 101 provides instructions 105 to GPU 102 for execution on behalf of program 103 , which in turn provides instructions for execution on behalf of program 104 . 107 , followed by other instructions (not shown) for execution on behalf of the program 103 . As further described herein, in some cases different programs, and thus different instructions, may be configured with different required clock frequencies and/or different required clock frequencies in order for the programs to meet other requirements, such as a specified quality or a specified display frame rate. same different specified processing requirements. GPU 102 identifies different processing requirements by analyzing the workloads generated based on the instructions, and adjusts processing parameters, such as clock frequency, based on the identified processing requirements. GPU 102 thereby dynamically adjusts processing parameters based on a combination of processing requirements and specified processing requirements, rather than based on fixed processing parameter values.

To facilitate execution of CPU instructions (eg, instructions 105 and 107 ), GPU 102 includes a scheduler 106 and a set of computing units 115 . The set of computing units 115 includes a plurality of computing units (eg, computing unit 118 ), each computing unit comprising a number of processing elements configured to perform graphics and vector processing operations; It includes, for example, a single instruction, multiple data (SIMD) units, stream processors, and the like. In some embodiments, each computing unit includes additional modules to support processing elements such as one or more branch units, scalar units, vector units, register files, and the like. The scheduler 106 is a module that schedules operations in the form of wavefronts in the set of computing units 115 based on instructions received from one or more CPUs. In some embodiments, GPU 102 includes an instruction processor or other module to decode received instructions into one or more operations and provides the operations to scheduler 106 for scheduling. Also, in some embodiments, scheduler 106 is configured with different scheduling modules, each scheduling module scheduling operations on different resources of GPU 102 . For example, in some embodiments, the scheduler 106 is a scheduling module for scheduling graphics and vector processing operations on the computing units 115 , a scheduling module for scheduling memory operations on memory resources of the GPU 102 . scheduling module and the like.

To synchronize operations in computing units 115 (as well as other modules), GPU 102 uses clock control module 110 to generate a clock signal CK and transmits the CK signal to the computing units. (115) provided for each. In some embodiments, the clock control module 110 includes one or more control loops, such as a frequency locked loop (FLL) for locking the frequency of the clock signal CK to a specified frequency. The frequency is adjustable via control signaling as further described herein. In particular, GPU 102 sets the frequency of clock signal CK based on processing requests placed on GPU 102 by programs 103 and 104 so that each program conforms to its processing specification. To set the control signaling for the clock control module (110).

To illustrate, the overall use of resources of GPU 102 based on a received instruction or set of instructions is referred to herein as a workload (eg, workloads 116 and 117 ). Heavier or higher workloads use more resources of GPU 102 , while lighter or lower workloads use fewer resources of GPU 102 . Thus, the workload generated by a particular one of the programs 103 and 104 is based on the instructions generated by the program. Also, the workload generated by a program generally correlates with the processing specifications for that program. Thus, for example, a program with a relatively high frequency of execution (ie, a program that is expected or specified to run quickly) generally creates a heavier workload (ie, a workload that requires more resources to execute). On the other hand, a program with a relatively low specified execution frequency creates a lighter workload (ie, a workload that requires more resources to run).

To accommodate concurrently executing programs having different specified execution frequencies, GPU 102 includes a clock frequency adjustment module (CFAM) 108 . The CFAM 108 monitors parameters indicative of the current workload of the GPU 102 , thereby in effect detecting which of the programs 103 and 104 are currently running on the GPU 102 , and detecting the detected Provides control signaling to the clock control module 110 to set the frequency of the clock signal CK to the specified clock frequency for the program. Examples of parameters monitored by the CFAM 108 are, in some embodiments, the number of wavefronts scheduled by the scheduler 106 at the computing units 115 at the specified amount of time, the specified amount of time. the number of draw instructions received by GPU 102 in , the type of draw or dispatch instruction, a hint provided by the compiler, etc., or any combination thereof. If the monitored parameters exceed the workload threshold, the CFAM 108 increases the frequency of the CK signal to a higher specific frequency F ₂ . In response to the monitored parameters falling below the workload threshold for a specified amount of time, the CFAM 108 reduces the clock frequency of the CK signal to a lower specified frequency F ₁ . In some embodiments, higher and lower specific frequencies F ₂ and F ₁ are applied to programs 103 and 104 during initialization of the corresponding program, such as through instructions provided to GPU 102 by each program. ) is indicated by Also, in some embodiments, the workload thresholds and frequencies F ₁ and F ₂ are programmable values that allow the programmer to tune the performance of the running programs to a desired level.

To illustrate by way of example, in some embodiments, the parameter monitored by CFAM 108 is the number of draw instructions received from each of programs 103 and 104 and scheduled to be executed by scheduler 106 . In response to a number of draw instructions during a specified amount of time exceeding a specified workload threshold (eg, 10 or more draw instructions over 100 execution cycles), CFAM 108 may Assume that a program associated with a specified clock frequency is running and requires a high number of resources. In response, the CFAM 108 increases the frequency of the CK signal to a higher specific frequency F ₂ . When the number of draw commands received falls below the specified workload threshold, the CFAM 108 assumes that the program associated with the lower specified frequency is being executed. In response, the CFAM 108 reduces the frequency of the CK signal to a lower specific frequency F ₁ .

As another example, in some embodiments, the parameter monitored by the CFAM 108 is the number of wavefronts scheduled to be executed on the computing units 115 by the scheduler 106 . In response to a number of scheduled wavefronts exceeding a workload threshold (eg, more than 100 wavefronts over 500 execution cycles), the CFAM 108 may initiate a program associated with a higher specified clock frequency. Assuming this is running, it requires a high number of resources. In response, the CFAM 108 increases the frequency of the CK signal to a higher specific frequency F ₂ . When the number of scheduled wavefronts falls below a specified workload threshold, CFAM 108 assumes that a program associated with a lower specific frequency is running. In response, the CFAM 108 reduces the frequency of the CK signal to a lower specific frequency F ₁ .

In other embodiments, the parameter monitored by the CFAM 108 is the number of draw instructions for drawing the specified object, the number of draw instructions of a specified type, such as an object having a threshold number of vertices, and the like. CFAM 108 determines the type of draw command from the command parameters included in commands 105 and 107 . For example, in some embodiments, each command indicates the type of object to be drawn, the number of vertices of the object, and the like. In other embodiments, these parameters are identified by the instruction processor of GPU 102 .

As indicated by the examples above, in some embodiments, the CFAM 108 identifies a workload of the GPU based on information stored or monitored in the scheduler 106 . For example, in some embodiments, scheduler 106 may control the number of draw instructions received from CPU 101 , types of draw instructions received, number of wavefronts scheduled for execution, etc., or any thereof maintain registers or other memory structures that store data representing a combination of Based on the stored information, the CFAM 108 identifies the overall workload for the GPU 102 and adjusts the clock frequency of the CK clock signal as described herein.

By changing the clock frequency based on the detected workload, rather than relying solely on static conditions such as program identifier values or specified time periods, GPU 102 can dynamically meet the processing specifications for each program and , thereby saving processor resources. For example, in some cases, a high-performance program may present a relatively light workload to GPU 102 (eg, to execute a relatively simple draw instruction), so that GPU 102 The frequency of the CK signal is maintained at a lower frequency (F ₁ ). In contrast, if GPU 102 only changed the clock frequency by a static condition such as a program identifier, the frequency of the CK signal would be increased to a higher frequency without a corresponding performance benefit.

2 illustrates a diagram 200 illustrating an example of a CFAM 108 that adjusts the frequency of a clock signal CK based on a detected workload in accordance with some embodiments. The diagram 200 shows an x-axis representing time and a y-axis representing the frequency of the clock signal CK. The diagram 200 further shows a plot 201 representing an example of the frequency of the clock signal CK varying over time.

In the example of FIG. 2 , it is assumed that program 103 is associated with a lower specific frequency designated F ₁ and program 104 is associated with a higher specific frequency. In the illustrated example of plot 201 , at initial time 202 , the workload of GPU 102 is below a workload threshold as indicated by one or more workload parameters. This indicates that GPU 102 is likely to execute instructions on behalf of program 103 . Accordingly, and in response to the workload being below the workload threshold, the CFAM 108 sets the frequency of the clock signal CK to a lower frequency F ₁ .

At time 203 after time 202 , the workload on GPU 102 increases, such that the workload exceeds the workload threshold. Thus, the workload indicates that GPU 102 is executing instructions on behalf of program 104 . Thus, and in response to the increasing workload beyond the workload threshold, the CFAM 108 controls the frequency of the clock signal CK until at time 204 a specific frequency F ₂ is reached with a higher frequency. starts to increase As illustrated, the CFAM 108 converts the clock signal from frequency F ₁ to frequency over time (between times 203 and 204 ) rather than immediately setting the clock frequency to F ₂ at time 203 . Ramp with (F2). In some embodiments, the time between time 203 and time 204 is 50 microseconds or less. By ramping the clock from frequency F ₁ to frequency F ₂ , GPU 102 stops executing and operates between times 203 and 204 rather than flushing data from computing units 115 . keep running them

Between time 204 and time 205 , the workload of GPU 102 exceeds the workload threshold, and in response CFAM 108 increases the frequency of clock signal CK to a higher frequency F _{2 .} ) to keep At time 205 , the workload of GPU 102 drops below the threshold. In response, the CFAM 108 begins to ramp the frequency of the clock signal CK to return to the lower frequency F ₁ . In some embodiments, the CFAM 108 uses hysteresis to prevent short excursions around the workload threshold from causing frequent adjustments to the frequency of the clock signal CK. For example, in some embodiments, the CFAM 108 initiates adjustment of the clock signal frequency in response to the workload of the GPU 102 being above a threshold for a specified amount of time.

3 shows an example of a CFAM 108 in accordance with some embodiments. In the illustrated example, CFAM 108 includes a control module 320 , a set of program frequency registers 322 , and a set of workload threshold registers 324 . Program frequency registers 322 are a set of programmable registers that store frequency values for each program executed in the processing system of GPU 102 . In some embodiments, each executing program sends instructions to GPU 102 via a device driver. When a program begins execution in the processing system, the program sends device driver frequency information indicating the specified execution frequency for the program. In response, the device driver programs a corresponding one of the program frequency registers 322 with the execution frequency specified for the program.

The workload threshold registers 324 are a set of programmable registers that store workload thresholds for each program executing in the processing system of the GPU 102 . In some embodiments, each executable program is associated with a workload profile, which indicates possible workloads generated by the program. In some embodiments, the workload profile is created by the software developer during development of the program. In other embodiments, the workload profile is developed by GPU 102 during the first N times the program is executed, where N is an integer. For example, the first N times that the program is executed, the GPU 102 determines the number of wavefronts scheduled for execution on behalf of the program, the number of draw instructions issued by the CPU 101 on behalf of the program, etc. Similarly, a performance monitor (not shown) is used to measure the workload generated by the program. The control module 320 generates a workload threshold indicative of the average workload of the program (eg, the average number of wavefronts generated by the program or the average number of draw instructions, or a combination thereof), Store the workload threshold in a corresponding one of the load threshold registers 324 .

In operation, when at least two programs are concurrently executed in the processing system of GPU 102 , control module 320 controls, from scheduler 106 , the number of scheduled wavefronts or draw commands received at a specified amount of time. Receive information such as numbers. The control module 320 compares the workload information with a workload threshold stored in the workload threshold register 324 . In response to the workload threshold being exceeded, the control module 320 determines the program associated with the exceeded threshold and retrieves the program frequency for the program from the program frequency register 322 . Then, the control module 320 transmits a control signaling to the clock control module 110 to adjust the frequency of the CK clock signal to the retrieved program frequency.

In response to the workload information indicating that the GPU workload has fallen below the workload threshold, the control module 320 executes the program associated with the next-lower workload threshold stored in the workload threshold registers 324 . decide The control module 320 retrieves the program frequency for the identified program from the program frequency registers 322 and sends a control signaling to the clock control module 110 to adjust the frequency of the CK clock signal to the retrieved program frequency. .

4 shows a flow diagram of a method 400 for setting the frequency of a clock signal in a processing unit based on identifying an executable program as indicated by a detected workload in accordance with some embodiments. For purposes of explanation, method 400 is described with respect to an example implementation in GPU 102 of FIG. 1 , although in other embodiments method 400 may be used in other processing units and other processing systems. It will be appreciated that the implementation Returning to the flowchart, at block 402 , GPU 102 determines a workload threshold for each program being executed in the processing system. As noted above, in some embodiments, each executable program provides a workload threshold upon initialization based on a workload profile generated during development of the program. In other embodiments, GPU 102 identifies a workload threshold for an executing program by determining an average workload generated by the program during the first N times the program is executed. In still other embodiments, GPU 102 records the workload generated by the running program each time the program is executed, and determines the average workload generated by the program during the last M times the program was executed. and determine a load threshold, where M is an integer. GPU 102 writes the workload threshold in a corresponding register of workload threshold registers 324 .

At block 404 , GPU 102 determines specified clock frequencies for concurrently executing programs in the processing system. As indicated above, in some embodiments, the specified clock frequency is provided to GPU 102 by the respective executable program via a device driver. GPU 102 stores the specified clock frequencies in corresponding ones of program frequency registers 322 .

At block 406 , the CFAM 108 executes the scheduler 106 , such as the number of wavefronts scheduled at the computing units 115 , the number of draw instructions received by the GPU 102 within a specified amount of time, and the like. monitor the workload of GPU 102 based on the information provided by At block 408 , the CFAM 108 determines whether the workload has exceeded one of the workload thresholds stored in the workload registers 324 . Otherwise, the method returns to block 406 where the CFAM 108 maintains the clock rate of the CK clock signal at the current frequency.

At block 408 , if the CFAM 108 determines that the workload has exceeded the workload threshold, then the CFAM 108 identifies the program associated with the exceeded threshold, as stored in the program frequency registers 322 . together further determine the specified program frequency for the program. At block 410 , the CFAM 108 sends control signaling to the clock control module 110 to adjust the frequency of the CK clock signal to the specified program frequency.

The method flow moves to block 412 and the CFAM 108 continues to monitor the workload of the GPU 102 based on the information provided by the scheduler 106 . At block 414 , the CFAM 108 determines whether the workload has fallen below a workload threshold stored in the workload register 324 . Otherwise, the method returns to block 412 and the CFAM 108 maintains the clock rate of the CK clock signal at the current frequency. In response to the workload falling below the workload threshold, the method flow moves to block 416 , where the CFAM 108 returns the frequency of the CK clock signal to the initial lower frequency of the clock control module 110 . ) to transmit control signaling. The method flow returns to block 406 , where the CFAM 108 continues to monitor the workload of the GPU 102 .

In some embodiments, the method comprises: receiving, at a graphics processing unit (GPU), a plurality of instructions from a central processing unit (CPU), the plurality of instructions being associated with a plurality of program threads executing concurrently on the CPU determining, at the GPU, a first workload to be executed on the GPU based on at least one of the plurality of instructions, each of the plurality of threads being associated with a corresponding specified clock frequency; identifying, based on the first workload, a first program thread from among the plurality of program threads that are concurrently executed in the CPU; and in response to identifying the first program thread, adjusting a clock signal of the GPU to the particular clock frequency associated with the first program thread. In an aspect, identifying the first program thread comprises identifying the first program thread in response to the first workload exceeding a first workload threshold.

In an aspect, a method further comprises: determining, at the GPU, a second workload to be executed on the GPU after the first workload based on at least one other of the plurality of instructions; identifying, based on the second workload, a second program thread from among the plurality of program threads that are concurrently executed in the CPU; and in response to identifying the second program thread, adjusting the clock signal of the GPU from the first frequency to the particular frequency associated with the second program thread. In another aspect, identifying the second program thread includes identifying the second program thread in response to the second workload being below a second workload threshold. In another aspect, the first threshold is programmable. In another aspect, identifying the first workload includes identifying the first workload based on information received from a scheduler of the GPU.

In an aspect, identifying the first workload includes identifying the first workload based on a number of wavefronts scheduled for execution on the set of computing units of the GPU. In another aspect, identifying the first workload includes identifying the first workload based on a type of draw command received at the GPU. In another aspect, adjusting the clock includes ramping the clock from the second frequency to the first frequency.

In some embodiments, the method includes: identifying, based on a first workload to be executed on a graphics processing unit (GPU), a first program thread among the plurality of program threads, wherein the GPU performs a work on behalf of the plurality of program threads. execute loads, and multiple program threads are executed concurrently in a central processing unit (CPU); and in response to identifying the first program thread, adjusting a clock of the GPU to a first frequency associated with the first program thread. In an aspect, identifying the first program thread includes identifying the first program thread in response to the first workload exceeding a first workload threshold. In another aspect, a method includes identifying a second program thread of a plurality of program threads based on a first workload to be executed on the GPU; and in response to identifying the second program thread, adjusting a clock of the GPU to a second frequency associated with the second program thread. In another aspect, adjusting the clock includes ramping the clock from the second frequency to the first frequency.

In some embodiments, the graphics processing unit (GPU) is a scheduler that receives a plurality of instructions from a central processing unit (CPU), the plurality of instructions being associated with a plurality of program threads executing concurrently on the CPU, each of the plurality of threads comprising: associated with a corresponding specific clock frequency -; a plurality of computing units configured to execute workloads based on the plurality of instructions; a clock control module for generating a first clock signal for the plurality of computing units; and the clock frequency adjustment module determines a first workload to be executed on the plurality of computing units; identify, based on the first workload, a first program thread from among the plurality of program threads that are concurrently executed in the CPU; and in response to identifying the first program thread, adjust the clock signal to a particular clock frequency associated with the first program thread. In an aspect, the clock frequency adjustment module is configured to identify the first program thread in response to the first workload exceeding the first workload threshold.

In an aspect, the clock frequency adjustment module is configured to: determine a second workload to be executed on the plurality of computing units; identify a second program thread based on the second workload; and in response to identifying the second program thread, adjust the clock signal from the first frequency to a second frequency associated with the second program thread. In another aspect, the clock frequency adjustment module is configured to identify the second program thread in response to the second workload being below the second workload threshold. In another aspect, the first threshold is programmable. In another aspect, the clock frequency adjustment module is configured to identify the first workload based on a number of scheduled wavefronts in the set of computing units of the GPU. In another aspect, the clock frequency adjustment module is configured to identify the first workload based on a number of draw commands received at the scheduler.

Computer-readable storage media may include any non-transitory storage media or combination of non-transitory storage media that are accessible by a computer system during use to provide instructions and/or data to the computer system. Such storage media may include optical media (eg, compact disc (CD), digital versatile disc (DVD), Blu-ray disc), magnetic media (eg, floppy disk, magnetic tape, or magnetic hard drive), volatile memory (eg, random access). memory (RAM) or cache), non-volatile memory (eg, read-only memory (ROM) or flash memory), or microelectromechanical system (MEMS) based storage media. A computer-readable storage medium may be embodied in a computing system (eg, system RAM or ROM), and may be fixedly attached to a computing system (eg, a magnetic hard drive) or a computing system (eg, an optical disk or It is connected to a computer system through a Universal Serial Bus (USB)-based flash memory), or a wired or wireless network (such as a Network Accessible Storage (NAS)).

In some embodiments, certain aspects of the techniques described above may be implemented by one or more processors of a processing system executing software. Software includes sets of one or more executable instructions stored on or otherwise tangibly embodied on a non-transitory computer-readable storage medium. Software may include specific data and instructions that, when executed by one or more processors, operate the one or more processors to perform one or more aspects of the techniques described above. The non-transitory computer-readable storage medium may include a magnetic or optical disk storage device, a flash memory, a cache, a solid state storage device such as RAM, or other non-volatile memory devices, and the like. The executable instructions stored on a non-transitory computer-readable storage medium may be in source code, assembly language code, object code, or other instruction format that is interpreted or otherwise executable by one or more processors.

Not all activities or elements described above in the general description are required, some of specific activities or devices may not be required, one or more additional activities in addition to those described above may be performed, or included elements are required. take note of Also, the order in which the activities are listed is not necessarily the order in which the activities are performed. In addition, the concept has been described with reference to specific examples. However, those of ordinary skill in the art can see that various modifications and changes are possible without departing from the scope of the present disclosure as described in the claims below. Accordingly, the specification and drawings are to be regarded in an illustrative rather than a restrictive sense, and all such modifications are intended to be included within the scope of the present disclosure.

Advantages, other advantages, and solutions to problems have been described above in connection with specific embodiments. However, an advantage, advantage, solution to a problem, and any feature from which any advantage, advantage, or solution may arise or become more pronounced shall be construed as an important, required, or essential feature of any one or all of the claims. doesn't happen Moreover, the specific embodiments disclosed above are exemplary only, and while the disclosed subject matter will be apparent to those skilled in the art having the benefit of the teachings herein, it can be modified and practiced in different but equivalent ways. No limitations are intended to the details of construction or design described herein other than as set forth in the claims below. Accordingly, it is apparent that the specific embodiments disclosed above may be altered or modified, and all such variations are considered within the scope of the disclosed subject matter. Accordingly, the scope of protection sought in this specification is as specified in the claims below.

Claims (20)

In the method,
receiving, at a graphics processing unit (GPU) 102 , a plurality of instructions ( 105 , 107 ) from a central processing unit (CPU) 101 , wherein the plurality of instructions are executed concurrently in the CPU by a plurality of program threads 103, 104, wherein each of the plurality of threads is associated with a corresponding particular clock frequency;
determining, at the GPU, a first workload (116) to be executed on the GPU based on at least one of the plurality of instructions;
identifying, based on the first workload, a first program thread from among the plurality of program threads that are concurrently executed in the CPU; and
in response to identifying the first program thread, adjusting a clock signal of the GPU to the particular clock frequency associated with the first program thread. According to claim 1,
wherein identifying the first program thread comprises identifying the first program thread in response to the first workload exceeding a first workload threshold. 3. The method of claim 2,
determining, at the GPU, a second workload (117) to be executed on the GPU after the first workload based on at least one other of the plurality of instructions;
identifying, based on the second workload, a second program thread from among the plurality of program threads that are concurrently executed in the CPU; and
in response to identifying the second program thread, adjusting the clock signal of the GPU from the first frequency to the particular frequency associated with the second program thread. 4. The method of claim 3,
wherein identifying the second program thread comprises identifying the second program thread in response to the second workload being below a second workload threshold. 3. The method of claim 2, wherein the first threshold is programmable. The method of claim 1 , wherein identifying the first workload comprises identifying the first workload based on information received at a scheduler (106) of the GPU. The method of claim 1 , wherein identifying the first workload comprises identifying the first workload based on a number of wavefronts scheduled for execution on a set of computing units of the GPU. Way. The method of claim 1 , wherein identifying the first workload comprises identifying the first workload based on a type of a draw command received at the GPU. 2. The method of claim 1, wherein adjusting the clock comprises ramping the clock from a second frequency to the first frequency. In the method,
identifying, based on a first workload (116) to be executed on a graphics processing unit (GPU) (102), a first program thread (103) of a plurality of program threads (103, 104) - the above GPU executes workloads on behalf of the plurality of program threads, wherein the plurality of program threads are concurrently executed in a central processing unit (CPU) 101 ; and
in response to identifying the first program thread, adjusting a clock of the GPU to a first frequency associated with the first program thread. 11. The method of claim 10,
wherein identifying the first program thread comprises identifying the first program thread in response to the first workload exceeding a first workload threshold. 12. The method of claim 11,
identifying a second program thread (104) of the plurality of program threads based on a first workload to be executed on the GPU; and
in response to identifying the second program thread, adjusting the clock of the GPU to a second frequency associated with the second program thread. 11. The method of claim 10, wherein adjusting the clock comprises ramping the clock from a second frequency to the first frequency. A graphics processing unit (GPU) (102) comprising:
a scheduler receiving a plurality of instructions (105, 107) from a central processing unit (CPU) (101), the plurality of instructions being associated with a plurality of program threads (104, 106) executing concurrently in the CPU, the plurality of each thread of − is associated with a corresponding specific clock frequency;
a plurality of computing units (115) configured to execute workloads (116, 117) based on the plurality of instructions;
a clock control module 110 for generating a first clock signal for the plurality of computing units; and
a clock frequency adjustment module (108), the clock frequency adjustment module comprising:
determine a first workload to be executed on the plurality of computing units;
identify, based on the first workload, a first program thread from among the plurality of program threads that are concurrently executed in the CPU; and
and in response to identifying the first program thread, adjust the clock signal to the particular clock frequency associated with the first program thread. 15. The method of claim 14, wherein the clock frequency adjustment module,
and identify the first program thread in response to the first workload exceeding a first workload threshold. 16. The method of claim 15, wherein the clock frequency adjustment module,
determine a second workload to be executed on the plurality of computing units;
identify a second program thread based on the second workload; and
and in response to identifying the second program thread, adjust the clock signal from a first frequency to a second frequency associated with the second program thread. The method of claim 16, wherein the clock frequency adjustment module,
and identify the second program thread in response to the second workload being below a second workload threshold. 16. The GPU of claim 15, wherein the first threshold is programmable. The GPU of claim 14 , wherein the clock frequency adjustment module is configured to identify the first workload based on a number of scheduled wavefronts in a set of computing units of the GPU. 15. The GPU of claim 14, wherein the clock frequency adjustment module is configured to identify the first workload based on a number of draw commands received at the scheduler.

Images